Skip to content
AI Coding

Goldman Pulls Claude Code: Use It Safely Yourself

Goldman Sachs is pulling Claude Code from bankers and traders. Here's what the risk review means and a beginner setup so you can use the agent without the same traps.

5 min readBeginner

Two reactions to Goldman Sachs pulling Claude Code from bankers and traders: copy the ban, or steal the lesson. A firm-wide cut stops unreviewed scripts on client data. Your laptop is not a trading floor – but silent billing, fat CLAUDE.md files, and “just allow Bash” still bite solo users the same way desk chaos did.

Bloomberg’s scoop (reported Nov 2, 2026) landed on October 2026 internal guidance: Claude Code and GitHub coding tools back to engineers, strats exempt. For months non-engineers had used the agent on flows, quick models, and content; July 2026 CIO Marco Argenti had praised vibe-coded prototypes handed to roughly 12,000 engineers. The in-house multi-model GS AI Assistant stayed firmwide. You still want the loop. Setup below is the personal version of their role gate.

What Claude Code actually is (and why banks flinch)

Plain-English goal in. Repo read, multi-file plan, edits, tests, shell, git out. That is Claude Code – Anthropic’s agentic coding system in terminal, desktop Code tab, VS Code/JetBrains, and web. Autocomplete it is not.

Banks flinch because the same loop that drafts a dashboard can touch production-shaped data with no ticket. Spokesperson Nick Carcaterra’s line (via Bloomberg): experiment, then put tools with people best suited long term after productivity and risk weigh-in. Their gate is job title. Yours is permission modes plus a short rules file.

Pro tip: New repo → stay Manual. Approve the first dozen actions by hand. Widen allow lists only for boring repeats. Muscle memory beats bypass.

Practical setup: install to first approved edit

Paid plan only – Free chat does not include Claude Code. As of late 2026 on the pricing page (re-check before you pay; this can change): Pro $20/mo ($17 annual), Max $100, Ultra $200. Chat and Code share one usage pool (rolling 5-hour window plus weekly cap).

# macOS / Linux / WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

claude --version
cd /path/to/your/project
claude

Same login as claude.ai. First prompt read-only: “what does this project do?” Files pull on demand – no zip upload.

  1. Run /init once for a starter CLAUDE.md.
  2. Edit it yourself. Under two screens. Hard rules at the top, imperative: “Never commit secrets. Always run pytest before claiming tests pass. IDs are strings.” Long soft prose gets skimmed; docs and community both flag that.
  3. Tiny change: “add a hello-world helper and show the diff.” Yes only if the diff is clean. Reject and say why when it isn’t.
  4. Shift+Tab cycles permission modes – only after you know which mode you are in. Manual stays the beginner default.

Full walkthrough: Claude Code quickstart. Desktop: paid sign-in → Code tab, skip the CLI if you want.

Advanced usage without the Goldman failure mode

Map first, edit second. “Outline the auth module and a three-step plan” before any multi-file rewrite. Git the same way: “what changed?” then “commit with a descriptive message” – not vague “ship it.”

claude "fix the failing test in auth and rerun until green"
claude -p "explain this function" # one-shot, exit
claude -c # continue last session in this dir

CLAUDE.md orients new sessions. It is memory, not a sandbox. Permissions and hooks enforce; prose only steers.

Same rule banks needed: draft with the agent, human owns merge and deploy when real data or credentials are in play. That split is the strats-shaped exception on a laptop.

Honest limitations and the gotchas tutorials skip

Default posture: ask before most writes and shell. Modes exist that loosen it.

bypassPermissions can skip prompts – including on protected paths like .git. Official permissions docs: isolated containers or VMs only. Broad Bash(*) allow rules are the twin foot-gun.

The catch is billing. If ANTHROPIC_API_KEY sits in your environment, Claude Code may prefer the key over Pro/Max/Ultra login and meter API rates instead of plan usage (Help Center plan-connection FAQ). Unset it, restart, confirm with /status when the meter looks wrong.

Quota: one pool across web/desktop/mobile and Code. A long chat thread can rate-limit you mid-refactor. Near a limit – lighter model, /clear between unrelated tasks, or wait out the rolling window. Durable rules live in CLAUDE.md, not the scrollback.

Goldman never published the productivity spreadsheet or a single smoking-gun incident. That gap is fine. You do not need their numbers to keep a human in the loop on sensitive workflows.

FAQ

Did Goldman ban all AI?

No. Coding agents went engineers-only (strats exempt). The multi-model in-house assistant stayed firmwide.

I’m not an engineer. Can I still use Claude Code productively?

Yes – Manual mode, short top-of-file rules, no unreviewed scripts on production data or credentials. Marketer cleaning a CSV copy: fine. Live client flow into an unreviewed script: the exact pattern desks are walking back. Describe the outcome, approve the diff, run tests yourself.

Pro vs Max for Claude Code beginners?

Pro for part-time learning. Max or Ultra when you live in long agent sessions and keep hitting shared limits with normal Claude chat. Pick from how hard you push both surfaces – not a magic message count. Prices move; use the live pricing page.

Throwaway folder tonight. Install for your OS. Paid sign-in. /init. Make it explain the empty project before it writes a line. That fifteen-minute drill beats waiting for the next policy email.

Related Tutorials